execution_plan.md

Day 1: Ingestion & Structuring (The Foundation)

• Goal: Turn a PDF into a clean, multi-modal dataset with UUIDs.

• Tasks:

  • Set up a Python script with Docling to parse the provided test PDFs.

  • Write a function to intercept any extracted Image or Table, send it to the VLM API, and get a text description.

  • Format the final output into a flat list of dictionaries, where every element has a chunk_id, type (text/table/image), content, and parent_header_id.

• Checkpoint: You can print a clean JSON payload representing the entire document layout.

Day 2: Hybrid Memory & Graph Mapping (The Brain)

• Goal: Build the search and relationship engine.

• Tasks:

  • Initialize ChromaDB. Embed the content of every chunk and store it with its chunk_id in the metadata.

  • Initialize a NetworkX directed graph.

  • Write the logic to iterate through your parsed JSON and add nodes and edges (e.g., Header_1 $\rightarrow$ Paragraph_A, Paragraph_A $\rightarrow$ Table_1).

  • Wrap this ingestion process in a FastAPI /upload endpoint that stores the resulting NetworkX graph in app.state.

• Checkpoint: You can query ChromaDB for a keyword, get a chunk_id, and successfully print its neighboring nodes from NetworkX.

Day 3: Agentic UI & Demo Polish (The Wow Factor)

• Goal: Connect the brain to the user and make it fast.

• Tasks:

  • Build the LangGraph agent: Query -> Vector DB (Find ID) -> NetworkX (Expand Context up to depth=1) -> LLM Prompt -> Answer.

  • Implement StreamingResponse in FastAPI so the answer types out instantly.

  • Build the React frontend. Implement the chat window on the left, and use React Flow on the right to visualize the nodes the AI traversed.

• Checkpoint: End-to-end testing with sample Dell documents. Practice the pitch.

Day 1: Ingestion & Structuring (The Foundation)

• Goal: Turn a PDF into a clean, multi-modal dataset with UUIDs.

• Tasks:

  • Set up a Python script with Docling to parse the provided test PDFs.

  • Write a function to intercept any extracted Image or Table, send it to the VLM API, and get a text description.

  • Format the final output into a flat list of dictionaries, where every element has a chunk_id, type (text/table/image), content, and parent_header_id.

• Checkpoint: You can print a clean JSON payload representing the entire document layout.

Day 2: Hybrid Memory & Graph Mapping (The Brain)

• Goal: Build the search and relationship engine.

• Tasks:

  • Initialize ChromaDB. Embed the content of every chunk and store it with its chunk_id in the metadata.

  • Initialize a NetworkX directed graph.

  • Write the logic to iterate through your parsed JSON and add nodes and edges (e.g., Header_1 $\rightarrow$ Paragraph_A, Paragraph_A $\rightarrow$ Table_1).

  • Wrap this ingestion process in a FastAPI /upload endpoint that stores the resulting NetworkX graph in app.state.

• Checkpoint: You can query ChromaDB for a keyword, get a chunk_id, and successfully print its neighboring nodes from NetworkX.

Day 3: Agentic UI & Demo Polish (The Wow Factor)

• Goal: Connect the brain to the user and make it fast.

• Tasks:

  • Build the LangGraph agent: Query -> Vector DB (Find ID) -> NetworkX (Expand Context up to depth=1) -> LLM Prompt -> Answer.

  • Implement StreamingResponse in FastAPI so the answer types out instantly.

  • Build the React frontend. Implement the chat window on the left, and use React Flow on the right to visualize the nodes the AI traversed.

• Checkpoint: End-to-end testing with sample Dell documents. Practice the pitch.